open access

https://hdl.handle.net/2115/100455

Reliability-Based Natural Language Interface for Fire Safety Simulation : A Local LLM Approach to CFAST Input Generation

英語

CFAST

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Creating accurate input decks for computational fire simulations is a time-intensive, error-prone bottleneck in performance-based design. To address this, a reliability-based natural-language interface was developed to generate CFAST input decks using a local large language model combined with retrieval-augmented generation, automated validation, and semantic correction. Evaluation across 150 diverse, held-out prompts demonstrates an end-to-end system reliability of 67.2% (95% CI: 62.1%–72.3%). Performance exhibited a strong inverse relationship with scenario complexity, achieving 76.5% reliability in simple one- to two-room cases, compared with 48.8% in multi-floor, deviceheavy configurations. System uncertainty was formally decomposed into syntax uncertainty (21.9% invalid JSON generation), executability failures, and semantic uncertainty (10.3% prompt mismatches among executable decks). A failure mode and effects analysis identified the dominant error sources as context truncation, missing compartments, and vent connectivity issues. Based on these quantified failure boundaries, a tiered deployment policy is proposed that maps scenario complexity to required levels of human review. Ultimately, this local, offline architecture ensures privacy and provides a verifiable framework for adopting generative AI in safety-critical engineering applications.

会議発表論文

Version of Record